Exploratory Inference Learning for Scribble Supervised Semantic Segmentation

Scribble supervised semantic segmentation has achieved great advances in pseudo label exploitation, yet suffers insufficient exploration for the mass of unannotated regions. In this work, we propose a novel exploratory inference learning (EIL) framework, which facilitates efficient probing on unlabeled pixels and promotes selecting confident candidates boosting evolved segmentation. The regions is formulated as an iterative decision-making process, where policy searcher learns to infer unknown space reward based contrastive measurement candidates. particular, devise with intra-class attraction inter-class repulsion feature w.r.t labels. labeled exploitation are jointly balanced improve segmentation, framed close-looping end-to-end network. Comprehensive evaluations benchmark datasets (PASCAL VOC 2012 PASCAL Context) demonstrate superiority our proposed EIL when compared other state-of-the-art methods scribble-supervised problem.